68. effect of varying embedding energy and energy wise ... 6/vol6issue01/ijcsit2015060168.pdf ·...

Effect of Varying Embedding Energy and Energy wise Sorting on Watermarking using Wavelet

Transforms of Orthogonal Transforms DCT, Walsh and Haar

Dr. H. B. Kekre#1, Dr. Tanuja Sarode*2, Shachi Natu+3 #Senior Professor, Computer Engineering Department, MPSTME,

NMIMS University *Associate Professor, TSEC, Computer Engg. Department, Mumbai University

+Ph. D. Research Scholar, Computer Engineering Department, MPSTME,

NMIMS University

Abstract— This paper proposes a robust watermarking technique using wavelets of well-known transforms DCT, Walsh and Haar. HL and LH bands are separately selected for watermark insertion. While inserting watermark, its energy is varied with 40% margin to original energy of region of host selected for insertion of watermark to study effect on robustness. Performance of proposed technique is evaluated against cropping, compression, noise addition, image resizing and histogram equalization attacks and compared for varying embedding energy. Increasing embedding energy leads to improved robustness as well as sometimes marginal improvement in imperceptibility when various attacks are performed on watermarked image. This performance is also found to be better than use of orthogonal transforms DCT, Walsh and Haar as proposed in our previous work and use of wavelets without energy-wise sorting of middle frequency coefficients.

Keywords— Watermarking, DCT, Walsh, Haar, wavelet transforms.

I. INTRODUCTION

Copyright protection of digital contents is a major issue faced due to proliferation of internet technology for their transmission. Apart from this, many tools are available for copying or altering the digital contents. Many solutions have been found in literature to avoid copyright abuse of digital contents. Watermarking of digital contents is one of the popular solutions. Some information of owner known as watermark is inserted secretly in digital contents called as host. This watermark can be recovered and can be used to identify owner. Three important aspects of watermarking need to be taken care of for good watermarking. First is that perceptual quality of host should not reveal existence of hidden watermark i.e. imperceptibility. Second, watermark should not be removable from host except the owner and should resist common attacks performed on host. If watermark survives various attacks performed on host, it is known to be robust. Third, watermark should be extractable without need of

original host and watermark. In the proposed technique, a blind watermarking for digital images has been proposed using wavelet transforms obtained from popular orthogonal transforms DCT, Walsh and Haar. Watermark insertion is carried out in HL and LH region separately to compare their performance against various attacks. While inserting watermark, its energy is varied in the range of ±40% of HL and LH region of host to observe its effect on robustness. This performance is also compared with orthogonal transforms DCT, Walsh and Haar and with wavelets of the same without sorting the HL and LH region energy-wise. Organization of this paper as follows: Section II presents review of related work in digital image watermarking. Section III describes proposed method. Section IV shows results of proposed method against various attacks. Section V shows comparison with wavelets itself but without energy-wise sorting of HL and LH region while embedding. Section VI presents conclusion.

II. REVIEW OF RELATED WORK

Watermarking on digital images can be performed by directly changing the pixel values of image known as spatial domain watermarking. Spatial domain watermarking is easy to perform but is at high risk of altering watermark or getting it detected. On the other hand in transform domain watermarking image is first transformed in frequency domain by using appropriate transform and these transformed coefficients are altered appropriately to insert watermark. Transformed domain techniques are more robust than spatial domain techniques because after inserting watermark in frequency coefficients and taking inverse transform, watermark is irregularly spread all over the image. Thus it becomes difficult for an attacker to detect or modify watermark. in literature, watermarking using Discrete Cosine Transform [1-4], Discrete Fourier Transform [5-8], Discrete Wavelet Transform [9-12] are available. Singular Value Decomposition (SVD) is yet another popular transformation technique used in inserting watermark into host. In [13, 14],

H. B. Kekre et al, / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 6 (1) , 2015, 302-313

www.ijcsit.com 302

watermarking using SVD is proposed. From literature, it has been observed that instead of using any single transform when they are used together, give better robustness.

In [15] a multiband wavelet transformation based watermarking is proposed. Depending on number of passes different bands are selected for watermarking together. Third level decomposition using wavelets and HL and LH bands was observed to be good performance-wise. Watermarking using two dimensional Walsh coding is proposed in [16]. Watermark is coded using two dimensional Walsh functions and is then embedded in low frequency coefficients in DCT of host image. In [17], watermarking using DCT and DWT is proposed. DCT coefficients of watermark are inserted into high frequency wavelet coefficients of host. In [18], another joint DCT-DWT watermarking approach is proposed. A binary watermark is scrambled using Arnold transform and is embedded in 3-level wavelet transform of host. DCT of each DWT sub-band is computed and PN sequence of watermark is embedded in middle frequency coefficients of DCT block. This method is observed to be robust against image enhancement and noise addition attacks. A dual digital image watermarking is proposed by Maha Sharkas et. al in [19]. A watermarking technique incorporates two watermarks in a host image for improved protection and robustness. A watermark, in form of a PN sequence called the secondary watermark, is embedded in the wavelet domain of a primary watermark before being embedded in the host image. Singh, Rawat and Agarwal proposed yet another DCT-DWT watermarking technique for colour and grayscale images [20]. Host image is decomposed up to 3rd level of wavelet decomposition and passed through chaotic encryption process. Watermark is embedded in the form of DCT with special coefficient shifting algorithm.

III. PROPOSED METHOD

Method proposed in this paper is extension of our previous work in [21]. In proposed method instead of using traditional Haar wavelet or any other popular wavelet transform, wavelet transforms generated from DCT, Walsh and Haar are used to embed watermark into host. These wavelet transforms are obtained using Kekre’s algorithm to generate wavelet transforms [22]. For simulation, set of five color bitmap host images of size 256x256 and a color bitmap watermark of size 128x128 is used. These images are shown in Fig. 1.

(a)Lenna (b)Mandrill (c)peppers (d)face (e)puppy (f)NMIMSFig. 1 (a)-(e) Host images and (f) watermark image used for simulation

As proposed in [22], a wavelet transform matrix of size MNxMN can be generated using same component orthogonal transform matrix with different sizes, MxM and NxN respectively or by usig two different component transform matrices of size MxM and NxN respectively. Thus for example, 256x256 size DCT wavelet matrix can be generated using two DCT matrices of size 128x128 and 2x2 or 64x64 and 4x4 or 32x32 and 8x8 etc. for the proposed method

wavelet transform matrices of size 256x256 for host and of size 128x128 for watermark are required. So there are pairs of various possible combinations for host and watermark such as {(64, 4), (32, 4)} which means 256x256 size wavelet matrix of DCT/ Walsh/ Haar for host using component matrix of size 64x64 and 4x4 and 128x128 size wavelet matrix of DCT/ Walsh/ Haar for watermark using component matrix of size 32x32 and 4x4. Other possible pairs are {(64,4), (16,8)}, {(64,4), (8,16)}, {(64,4), (4,32)}, {(32,8), (32,4)}, {(32,8), (16,8)}, {(32,8), (8,32)}, … and so on. All such possible pairs for each wavelet transform have been tested and the pair which gives higher robustness for maximum number of attacks performed is selected. Steps for embedding are listed below.

1. Separate Red, Green and Blue channels of host image and watermark image and apply the selected pair of wavelet transform on them.

2. Sort the HL/LH band coefficients of host and all coefficients watermark in descending order of their energy values.

3. To bring the watermark coefficients in the uniform range of 0 to 1, they are normalized.

4. To bridge the energy gap between HL/LH band of host and watermark, watermark coefficients are scaled using suitable scaling factor such that energy of watermark is equal to the energy of HL/LH band of host.

5. Such normalized, scaled and sorted coefficients are now placed at the place of sorted coefficients of host.

6. Inverse wavelet transform is applied to transformed coefficients of host to get watermarked image.

To study the effect of varying embedding energy, scaling factor is adjusted such that it makes the energy of embedded watermark is less (60%) than that of HL/LH band of host image and greater (140%) than that of HL/LH band of host image.

For extraction of watermark, reverse procedure is followed. 1. Apply wavelet transform on watermarked image. 2. From HL/LH band extract watermark coefficients from

the positions where they were embedded. 3. Scale up the watermark coefficients using the same

scaling factor used in embedding process to get them in their original energy form followed by denormalization.

4. Take inverse wavelet transform to get the watermark. Fig. 2 below shows Lena host image after inserting

watermark and extracted watermark from it from HL and LH band before any attack is performed. Below each image corresponding MAE is displayed. Also the best size combination for which these results are obtained is given for HL and LH band.

Obtained watermarked images are subjected to various attacks and performance is evaluated for robustness. Description of these attacks and their results is given in following section.


www.ijcsit.com 303

Transform used for embedding watermark

Watermarked image

Extracted Watermark

Watermarked image

Extracted Watermark

HL band LH band

DCT wavelet

MAE 1.366 0 2.048 0 Size combination {(16,16),(16,8)} {(32,8),(32,4)}

Walsh wavelet


Haar wavelet


Fig. 2 Watermarked image Lena and extracted watermark from it using DCT wavelet, Walsh wavelet and Haar wavelet from HL and LH band when energy of watermark is matched (100%) with energy of HL and LH band of wavelet transformed host

IV. RESULTS OF PROPOSED METHOD AGAINST VARIOUS

ATTACKS:

A. Cropping of watermarked image

Watermarked images are cut at corners such that 16x16 size portion is removed from each corner. Variations to this attack are done in two ways, by increasing the cropped information from watermarked image and by keeping the amount of cropped information same but changing the position of cropped portion. To see the impact of increased amount of

cropping, 32x32 size portion is cropped at corners. to observe the effect of cropping other than at corners, but keeping the amount of cropped information same, 32x32 size portion is cropped at centre of watermarked image. Representative example of Lena image when subjected to 16x16 cropping at corners and watermark recovered from it is shown in Fig. 3. These results are for DCT wavelet, Walsh wavelet and Haar wavelet using HL and LH band for embedding.


Cropped Watermarked

image

Extracted Watermark

Cropped Watermarked

image

Extracted Watermark

HL band LH band

DCT wavelet

MAE 2.145 0.267 2.145 0.349

Walsh wavelet

MAE 2.145 0.690 2.145 5.765


www.ijcsit.com 304

Haar wavelet

MAE 2.145 0.542 2.145 1.219 Fig. 3 Result images for 16x16 cropping at corners when watermark is embedded into HL and LH band using DCT wavelet, Walsh wavelet and Haar

wavelet

Overall performance for cropping attack is shown in Fig. 4 graphically where average of MAE between embedded and extracted watermark from five different hosts is considered to judge the performance. For other attacks also average MAE between embedded and extracted watermark is considered.

Fig. 4 Comparison of DCT wavelet, Walsh wavelet and Haar wavelet

against cropping attack when HL and LH band is used to embed watermark

From Fig. 4 it is observed that embedding watermark in HL band leads to better robustness than using LH band for DCT

wavelet and Haar wavelet. This is applicable to Walsh wavelet also but only for 32x32 cropping at corners. For 16x16 cropping at corners and 32x32 cropping at centre, LH band used for embedding shows better robustness. B. Compression of watermarked image

The most obvious attack possible while the digital contents are exchanged over network is compression of data. For still images used in the simulation work, compression of images is performed in variety of ways namely compression using transformation techniques, compression using JPEG (quality factor 100%) and compression using vector quantization. For Vector quantization Kekre’s Fast Codebook Generation algorithm [23] is used and image is compressed using codebook of size 256. In compression attack using transforms, different orthogonal transforms like DCT, DST, Walsh, Haar and DCT wavelet are used to compress the watermarked image. Result images for Lena when subjected to JPEG compression and VQ compression are shown in Fig. 5 and Fig. 6 respectively.

Transform used for

embedding watermark

JPEG compressed Watermarked

image

Extracted Watermark

JPEG compressed Watermarked

image

Extracted Watermark

HL band LH band

DCT wavelet

MAE 2.055 67.141 1.881 58.10

Walsh wavelet

MAE 2.065 34.957 2.987 29.111

Haar wavelet

MAE 1.958 108.415 2.387 64.487 Fig. 5 watermarked image Lena after JPEG compression and extracted watermark from it when watermark is embedded in HL and LH band using DCT wavelet,

Walsh wavelet and Haar wavelet


www.ijcsit.com 305

From Fig. 5, it is observed that MAE between embedded and extracted watermark is quite high but it does not correspond to distortion of extracted watermark but due to

changes taking place in pixel intensity values. Also, as compared to HL band, LH band when used to embed watermark gives less MAE value for extracted watermark.

Transform used for

embedding watermark

VQ compressed Watermarked

image

Extracted Watermark

VQ compressed Watermarked

image

Extracted Watermark

HL band LH band

DCT wavelet

MAE 2.519 64.841 2.465 49.112

Walsh wavelet

MAE 2.54 48.693 2.792 24.961

Haar wavelet

MAE 2.493 103.823 2.650 48.656 Fig. 6 Watermarked image Lena when subjected to VQ based compression and watermark recovered from it using DCT wavelet, Walsh wavelet and Haar

wavelet for embedding watermark in HL and LH band of host.

From Fig. 6, it can be noted that for Lena image, MAE between embedded and extracted watermark is less when watermark is embedded in LH band instead of HL band for the wavelet transforms explored in proposed technique. Further Walsh wavelet with HL/LH band used for embedding

watermark proves more robust against compression using Vector Quantization than DCT wavelet and Haar wavelet.

Fig. 7 (a) shows the performance comparison of three wavelet transforms against compression using various transforms and Fig. 7 (b) shows the comparison against compression using DCT wavelet, JPEG compression and VQ.

(a) (b) Fig. 7 (a) Performance comparison of DCT wavelet, Walsh wavelet and Haar wavelet against compression using DCT, DST, Walsh and Haar when

watermark is embedded in HL and LH band (b) Performance comparison of DCT wavelet, Walsh wavelet and Haar wavelet against compression using DCT wavelet, JPEG and VQ when watermark is embedded in HL and LH band


www.ijcsit.com 306

C. Noise addition to watermarked images:

Another common attack watermarked image is addition of noise to it. In the proposed technique binary distributed random run length noise and Gaussian distributed run length noise is added to watermarked image. In case of binary

distributed run length noise run length is increased from range 1-10 to 5-50 and 10-100. Fig. 8 shows watermarked image Lena using various wavelet transforms after adding binary distributed run length noise with run length in the range of 10 to 100 and extracted watermark.


Watermarked image after adding Binary distributed

run length noise (10-100)

Extracted Watermark

Watermarked image after adding Binary distributed

run length noise (10-100)

Extracted Watermark

HL band LH band

DCT wavelet

MAE 1 8.797 1 0.255

Walsh wavelet

MAE 1 4.614 1 0

Haar wavelet

MAE 1 4.659 1 0

Fig. 8 Watermarked image Lena after adding binary distributed run length noise with run length 10-100 and watermark extracted from it when embedding is done HL and LH band using DCT wavelet, Walsh wavelet and Haar wavelet transform

Fig. 8 shows that embedding watermark in LH band is more robust than embedding in HL band against binary distributed run length noise of run length 10-100. Among the three wavelet transforms used, Walsh wavelet and Haar wavelet show highest robustness with their LH band used for embedding immediately followed by DCT wavelet.

Fig. 9 shows comparison of DCT wavelet, Walsh wavelet and Haar wavelet when their HL and LH bands are used to embed watermark against noise addition attack.

Fig. 9 comparison of wavelet transforms with their HL and LH bands used for

embedding watermark against noise addition attack (BRLN=Binary distributed run length noise and numbers in bracket specify run length,

GRLN=Gaussian distributed run length noise)

From Fig. 9 it is observed that, for binary run length noise of less run length i. e. 1 to 10, embedding watermark in HL band shows strong robustness with MAE between embedded and extracted watermark zero. For increased run length, embedding watermark in LH band shows better robustness. Also for increased run length, Haar wavelet consistently shows very good robustness closely followed by Walsh wavelet and then DCT wavelet in LH band. In HL band, for increased run length of binary distributed run length noise, Walsh wavelet is closely followed by Haar wavelet in robustness. For Gaussian distributed run length noise, performance of HL band of all three wavelets is exceptionally good when compared to LH band. In LH band, once again Walsh and Haar wavelets show equally well performance. D. Resizing of watermarked image:

In this type of attack, watermarked image is doubled in size and then reduced back to its original size. From such zoomed-reduced watermarked image, watermark is recovered. Different techniques used to perform zooming and reducing of watermarked images are, bicubic interpolation, image zooming using orthogonal transforms [24] and grid based


www.ijcsit.com 307

resizing [25]. Fig. 10 shows zoomed-reduced watermarked image using grid based resizing.


Zoomed reduced Watermarked

image using grid based resizing

Extracted Watermark

Zoomed reduced Watermarked

image using grid based resizing

Extracted Watermark

HL band LH band

DCT wavelet

MAE 0.250 4.171 0.291 52.448

Walsh wavelet

MAE 0.265 11.545 0.251 1.750

Haar wavelet

MAE 0.286 68.616 0.269 10.855 Fig. 10 Watermarked image Lena zoomed and reduced using Grid based resizing and extracted watermark when HL and LH bands are used to embed

watermark using DCT wavelet, Walsh wavelet and Haar wavelet

From Fig. 10 it can be said that for Lena image, when DCT wavelet is used to embed watermark, selection of HL band gives higher robustness than LH band. For Walsh wavelet and Haar wavelet, embedding watermark in LH band gives significantly better robustness over HL band. Among the three wavelet transforms, Walsh wavelet with LH band used for embedding is observed to be more robust against resizing attack. Performance of three wavelets in HL and LH band against resizing attack for five different host images is compared in the graph shown in Fig. 11.

Fig. 11 Comparison of DCT wavelet, Walsh wavelet and Haar wavelet

against resizing using bicubic interpolation, grid based interpolation and using DFT

From Fig. 11 following observations are made. For bicubic interpolation based resizing, DCT wavelet performs better than Walsh wavelet and Haar wavelet. Performance of DCT wavelet against resizing using bicubic interpolation is almost same in HL and LH band and for grid based resizing it is

more robust when embedding of watermark is done in HL band than in LH band. For grid based resizing, Walsh wavelet is more robust than DCT wavelet and Haar wavelet and when compared between HL and LH band, LH band shows better robustness. For DCT wavelet, embedding watermark in HL band and for Haar wavelet, embedding in LH band leads to robust watermarking against grid based resizing. For zooming-reducing using Discrete Fourier Transform (DFT), both HL and LH band selection for embedding watermark in all three wavelets proves to be exceptionally robust with minimum MAE. Moreover, for other transforms used for zooming-reducing watermarked images, gives zero MAE between embedded and extracted watermark.

V. PERFORMANCE COMPARISON WITH WAVELETS ITSELF BUT

WITHOUT ENERGY-WISE SORTING

Proposed technique is compared with our previous work wherein watermark is inserted in host image’s HL and LH band without sorting coefficients of host as well as watermark. A. Cropping attack

Table I-III show comparison of MAE values between embedded and extracted watermark obtained against cropping attack using sorting for embedding and without using sorting for embedding for DCT wavelet, Walsh wavelet and Haar wavelet. (Note: EE= Embedding Energy).

From Table I, it can be seen that for embedding in HL/LH band, by sorting of host and watermark coefficients, performance against cropping at corners is significantly


www.ijcsit.com 308

improved with reduced MAE values. Performance against cropping at centre is also improved in terms of robustness.

However, change in embedding energy does not change the MAE values between embedded and extracted watermark.

TABLE I COMPARISON OF DCT WAVELET AGAINST CROPPING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND USING

SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%)

DCT wavelet (HL) DCT wavelet (LH) 60% EE 100% EE 140% EE 60% EE 100% EE 140% EE

No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

16x16 crop

17.385 4.578 17.385 4.578 17.385 4.578 19.628 17.759 19.628 17.759 19.628 17.759

32x32 crop

36.393 7.959 36.393 7.959 36.393 7.959 38.480 29.478 38.480 29.478 38.480 29.478

32x32 crop

center 2.241 1.246 2.241 1.246 2.241 1.246 1.290 3.074 1.290 3.074 1.290 3.074

TABLE II COMPARISON OF WALSH WAVELET AGAINST CROPPING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND

USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%)

Walsh wavelet (HL) Walsh wavelet (LH) 60% EE 100% EE 140% EE 60% EE 100% EE 140% EE

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

16x16 crop

17.381 6.535 17.381 6.535 17.381 6.535 11.657 5.313 11.657 5.313 11.657 5.313

32x32 crop

36.099 11.928 36.099 11.928 36.099 11.928 36.738 16.379 36.738 16.379 36.738 16.379

32x32 crop

center 3.676 5.130 3.676 5.130 3.676 5.130 5.466 3.892 5.466 3.892 5.466 3.892

From Table II, we can conclude that for Walsh wavelet and HL/LH band used for embedding watermark, sorting of coefficients improves robustness against cropping at corners.

However, for cropping done at the centre of an image, MAE values are observed to be slightly increased irrespective of embedding energy.

TABLE III COMPARISON OF HAAR WAVELET AGAINST CROPPING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND

USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) Haar wavelet (HL) Haar wavelet (LH) 60% EE 100% EE 140% EE 60% EE 100% EE 140% EE

No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

16x16 crop

17.385 4.879 17.385 4.879 17.385 4.879 19.62783 6.858 19.62783 6.858 19.62783 6.858

32x32 crop

36.110 9.216 36.110 9.216 36.110 9.216 38.47501 33.627 38.47501 33.627 38.47501 33.627

32x32 crop

center 2.200 2.027 2.200 2.027 2.200 2.027 1.112732 2.566 1.112732 2.566 1.112732 2.566

Table III shows that when Haar wavelet and HL/LH band is used for embedding, sorting improves robustness against cropping image at corners. For cropping at centre, performance of sorting and without sorting is almost same for embedding in HL band. For LH band, performance of cropping at centre is slightly better when coefficients of host and watermark are not sorted.

B. Compression attack

Table IV-VI show comparison of MAE values between embedded and extracted watermark obtained against compression attack using sorting for embedding and without

using sorting for embedding for DCT wavelet, Walsh wavelet and Haar wavelet.

From Table IV, it can be seen that after sorting the transform coefficients during embedding in HL band, MAE between embedded and extracted watermark is reduced for compression using DCT, DST, Walsh, Haar and VQ for all variations of embedding energy. For compression using DCT wavelet, almost same performance is observed with and without sorting. For JPEG compression, sorting does not improve the robustness. Similar observations are noted for embedding in LH band also except that for DCT wavelet compression, significant improvement in robustness is noted.


www.ijcsit.com 309

TABLE IV COMPARISON OF DCT WAVELET AGAINST COMPRESSION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND

USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) DCT wavelet (HL) DCT wavelet (LH)

Compression Technique

60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

DCT 17.734 5.727 14.920 4.562 13.487 4.111 26.185 9.152 21.753 6.863 19.476 5.931 DST 18.275 5.869 15.161 4.644 13.587 4.162 25.509 9.334 21.006 6.946 18.712 5.965

Walsh 54.786 31.365 42.733 21.011 36.342 16.104 59.417 33.720 46.295 22.941 39.337 17.805 Haar 83.591 60.023 76.349 53.025 72.718 50.560 84.275 79.810 76.606 77.020 72.782 76.152

DCT wavelet 14.364 14.456 14.364 14.456 14.364 14.456 43.250 18.048 37.845 16.141 35.127 15.855 JPEG 43.197 58.400 41.961 61.996 41.497 64.907 44.921 50.279 42.892 57.444 42.180 63.111 VQ 69.696 59.958 56.478 55.300 49.150 53.557 71.806 62.612 57.720 56.783 50.480 53.682

TABLE V COMPARISON OF WALSH WAVELET AGAINST COMPRESSION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS

AND USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) Walsh wavelet (HL) Walsh wavelet (LH)


60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

DCT 42.285 42.450 33.460 40.769 28.843 40.203 42.490 44.874 35.449 44.341 30.394 44.278 DST 42.369 42.563 33.448 40.873 28.786 40.308 42.825 44.959 35.752 44.335 30.662 44.191

Walsh 6.298 39.665 6.298 39.665 6.298 39.665 4.181 49.346 4.181 49.346 4.181 49.346 Haar 60.512 56.521 60.512 56.521 60.512 56.521 59.208 72.183 59.208 72.183 59.208 72.183


TABLE VI COMPARISON OF HAAR WAVELET AGAINST COMPRESSION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND


Haar wavelet (HL) Haar wavelet (LH)


60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

DCT 48.579 33.035 41.182 29.126 37.468 27.651 49.486 29.670 40.329 25.225 35.613 23.756 DST 49.903 33.289 42.435 29.325 38.683 27.824 51.038 29.840 41.817 25.273 37.072 23.760

Walsh 38.292 29.270 38.292 29.269 38.292 29.269 39.406 29.063 39.406 29.063 39.406 29.063 Haar 55.878 56.426 55.878 56.426 55.878 56.426 56.360 73.981 56.360 73.981 56.360 73.981


From Table V, it is seen that Walsh wavelet leads to more

robustness after sorting only for JPEG and VQ compression attack in both HL and LH band and to some extent for Haar compression in HL band.

Table VI shows that when Haar wavelet is used to embed watermark either in HL or LH band, sorting leads to improved robustness for compression using DCT, DST, Walsh and DCT wavelet. For Walsh and Haar compression, increased embedding energy also improves robustness.

C. Noise addition attack

Table VII-IX show the comparison of embedding watermark with and without sorting transform coefficients of DCT wavelet transformed host and watermark when watermark is embedded in HL and LH band and different types of noises are added to it.

Table VII shows that sorting of host and watermark coefficients, improves robustness of noise addition attack. For smaller run length (1 to 10) of binary distributed run length noise, embedding in HL band with and without sorting is highly robust with zero MAE.

From Fig. VIII, we can conclude that, Walsh wavelet performs best against smaller run length of binary distributed run length noise attack and for Gaussian run length noise in HL band. Other run length noise when added row wise and column wise, robustness is improved. When watermark is embedded in LH band, proposed method shows strong robustness against increased run length of binary distributed run length.


www.ijcsit.com 310

TABLE VII COMPARISON OF DCT WAVELET AGAINST NOISE ADDITION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS

AND USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) DCT wavelet (HL) DCT wavelet (LH)

Noise type 60% EE 100% EE 140% EE 60% EE 100% EE 140% EE

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

BRLN (1- 10)

0.000 0.000 0.000 0.000 0.000 0.000 10.187 2.017 8.031 1.409 6.546 1.121

BRLN (5- 50)

29.926 10.801 23.541 6.993 20.176 5.412 2.995 0.958 2.352 0.712 2.023 0.605

BRLN (10- 100)

31.447 10.320 23.829 6.892 20.210 5.280 0.911 0.451 0.686 0.351 0.565 0.286

GRLN 0.733 0.277 0.568 0.216 0.480 0.185 33.992 12.464 26.338 8.280 22.261 6.308

TABLE VIII COMPARISON OF WALSH WAVELET AGAINST NOISE ADDITION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM

COEFFICIENTS AND USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) Walsh wavelet (HL) Walsh wavelet (LH)


No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

BRLN (1- 10)

0.000 0.000 0.000 0.000 0.000 0.000 7.824 2.229 7.663 1.385 6.287 1.083

BRLN (5- 50)

22.839 6.530 17.679 4.293 14.988 3.389 2.940 0.908 2.390 0.705 1.994 0.595

BRLN (10- 100)

23.340 6.419 17.714 4.348 15.002 3.242 0.000 0.000 0.000 0.000 0.000 0.000

GRLN 0.000 0.000 0.000 0.000 0.000 0.000 23.070 7.485 18.909 5.167 15.978 4.087

TABLE IX COMPARISON OF HAAR WAVELET AGAINST NOISE ADDITION ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS

AND USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%)

Haar wavelet (HL) Haar wavelet (LH)


No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

BRLN (1- 10)

0.000 0.000 0.000 0.000 0.000 0.000 8.235 2.027 5.740 1.228 4.625 0.838

BRLN (5- 50)

22.774 6.848 17.708 4.692 15.042 3.598 1.569 0.777 1.190 0.601 1.022 0.508

BRLN (10- 100)

23.283 6.294 17.842 4.349 15.151 3.565 0.000 0.000 0.000 0.000 0.000 0.000

GRLN 0.000 0.000 0.000 0.000 0.000 0.000 24.693 7.580 19.131 5.061 16.168 3.897

From Table IX, it is noted that using Haar wavelet and HL band shows strong robustness without sorting as well as with sorting for run length 1 to 10 and for Gaussian distributed run length noise. In LH band, higher run length (10-100) of binary distributed run length noise leads to strong robustness with and without sorting.

D. Resizing attack

Table X-XII show comparison of three wavelet transforms against resizing attacks when embedding is done with or without sorting in both in HL band and LH band.

TABLE X COMPARISON OF DCT WAVELET AGAINST RESIZING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND


DCT wavelet (HL) DCT wavelet (LH)


60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

Resize4 25.674 31.848 24.626 32.115 24.135 32.279 32.310 32.719 30.067 32.895 29.040 33.030 Resize 2 26.458 32.821 25.393 33.093 24.894 33.260 33.207 33.688 30.923 33.870 29.878 34.007

DFT_resize 4.479 0.707 3.494 0.580 2.971 0.527 4.437 0.985 3.450 0.780 2.927 0.682 Grid based 9.800 14.873 9.626 22.237 9.883 28.252 9.121 23.966 8.961 34.305 9.355 41.685

From Table X, it is observed that for DFT based resizing, sorting results in improved robustness. However, for bicubic

interpolation and grid based resizing, embedding without sorting is better than embedding with sorting. Another notable


www.ijcsit.com 311

thing is for transform based zooming other than DFT, MAE between embedded and extracted watermark is zero irrespective of embedding energy, frequency band used for embedding and transform used for embedding.

From Table XI, it is concluded that when embedding is done in HL band using Walsh wavelet, sorting does not improve robustness. But when used in LH band, for all types of resizing attack, sorting proves to be more robust.

TABLE XI COMPARISON OF WALSH WAVELET AGAINST RESIZING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND


DCT wavelet (HL) DCT wavelet (LH)


60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

Resize4 43.543 53.842 37.198 53.402 34.020 53.504 7.824 2.229 7.663 1.385 6.287 1.083 Resize 2 44.719 55.257 38.230 54.804 34.981 54.909 2.940 0.908 2.390 0.705 1.994 0.595

DFT_resize 0.489 1.906 0.489 1.906 0.489 1.906 0.000 0.000 0.000 0.000 0.000 0.000 Grid based 12.900 10.092 12.278 15.438 12.336 20.203 23.070 7.485 18.909 5.167 15.978 4.087

TABLE XII COMPARISON OF HAAR WAVELET AGAINST RESIZING ATTACK WHEN EMBEDDING IS DONE WITHOUT SORTING THE TRANSFORM COEFFICIENTS AND

USING SORTING IN HL AND LH BANDS FOR VARYING EMBEDDING ENERGY (60%, 100%, 140%) Haar wavelet (HL) Haar wavelet (LH)


60% EE 100% EE 140% EE 60% EE 100% EE 140% EE No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting No

sorting Sorting

No sorting

Sorting

Resize4 61.117 50.220 55.143 49.429 52.256 49.371 63.882 47.257 57.043 46.133 53.770 46.098 Resize 2 62.617 51.541 56.494 50.727 53.534 50.668 65.448 48.508 58.438 47.349 55.082 47.312

DFT-resize 9.588 1.552 9.588 1.552 9.588 1.552 6.725 1.621 6.725 1.621 6.725 1.621 Grid based 18.498 52.259 21.205 66.052 24.564 75.381 18.458 23.007 22.350 32.049 26.105 39.450

From Table XII, we can conclude that except grid based resizing, for all other resizing attacks, sorting leads to better robustness especially for DFT based resizing in HL and LH band.

VI. CONCLUSIONS

Conclusion regarding frequency band (HL/LH) of DCT wavelet, Walsh wavelet and Haar wavelet that shows better robustness against various attacks is shown in Table XIII.

TABLE XIII WAVELET TRANSFORMS OF DCT, WALSH AND HAAR WITH THEIR

HL/LH BAND SHOWING ROBUSTNESS AGAINST ATTACKS

Attack DCT

wavelet Walsh

wavelet Haar

wavelet Cropping HL HL, LH HL

Transform based compression

HL, LH HL HL

JPEG LH HL, LH LH VQ HL, LH HL, LH LH

BRLN 1-10, GRLN HL, LH HL, LH HL, LH BRLN 5-50, 10-100 LH LH LH Bicubic interpolation

resizing HL, LH HL LH

Grid based resizing HL LH LH

Apart from this, sorting improves the performance of all wavelet transforms tested in proposed technique against cropping and noise addition attack. Sorting improves the performance of Haar wavelet against resizing attack. Against compression attack, sorting leads to improved robustness of DCT wavelet and Haar wavelet to some extent. For majority of attacks except JPEG and VQ compression and grid based

resizing, increased embedding energy leads to better robustness.

REFERENCES [1] Cox, Ingemar J., Joe Kilian, F. Thomson Leighton, and Talal Shamoon.

"Secure spread spectrum watermarking for multimedia", IEEE Transactions on Image Processing, 6, no. 12 (1997): 1673-1687.

[2] Podilchuk, Christine I., and Wenjun Zeng. "Image-adaptive watermarking using visual models", IEEE Journal on Selected Areas in Communications, 16, no. 4 (1998): 525-539.

[3] Barni, Mauro, Franco Bartolini, Vito Cappellini, and Alessandro Piva. "A DCT-domain system for robust image watermarking", Signal processing 66, no. 3 (1998): 357-372.

[4] Gonge, Sudhanshu S., and Jagdish W. Bakal, "Robust digital watermarking technique by using DCT and spread spectrum", International Journal of Electrical, Electronics and Data Communication, 2320-2084.

[5] Tsui, Tsz Kin, Xiao-Ping Zhang, and Dimitrios Androutsos. "Color image watermarking using multidimensional Fourier transforms." Information Forensics and Security, IEEE Transactions on 3, no. 1 (2008): 16-28.

[6] Ramkumar, Mahalingam, Ali N. Akansu, and A. Aydin Alatan. "A robust data hiding scheme for images using DFT", In IEEE Proceedings of International Conference on Image Processing, vol. 2, pp. 211-215. 1999.

[7] Premaratne, P., and C. C. Ko, "A novel watermark embedding and detection scheme for images in DFT domain", In Seventh International Conference on Image Processing and Its Applications, IET, 1999. (Conf. Publ. No. 465), vol. 2, pp. 780-783.

[8] Kitamura, Itaru, Satoshi Kanai, and Takeshi Kishinami, "Copyright protection of vector map using digital watermarking method based on discrete Fourier transform", In Geoscience and Remote Sensing Symposium, 2001. IEEE 2001 International, vol. 3, pp. 1191-1193.

[9] Hsieh, Ming-Shing, Din-Chang Tseng, and Yong-Huai Huang, "Hiding digital watermarks using multiresolution wavelet transform", Industrial Electronics, IEEE Transactions on 48, no. 5 (2001): 875-882.

[10] Kim, Jong Ryul, and Young Shik Moon, "A robust wavelet-based digital watermarking using level-adaptive thresholding", In proceedings of IEEE International Conference on Image Processing, 1999, vol. 2, pp. 226-230.


www.ijcsit.com 312

[11] Lin, Wei-Hung, Shi-Jinn Horng, Tzong-Wann Kao, Pingzhi Fan, Cheng-Ling Lee, and Yi Pan. "An efficient watermarking method based on significant difference of wavelet coefficient quantization." Multimedia, IEEE Transactions on 10, no. 5 (2008): 746-757.

[12] Lin, Wei-Hung, Yuh-Rau Wang, Shi-Jinn Horng, Tzong-Wann Kao, and Yi Pan. "A blind watermarking method using maximum wavelet coefficient quantization", Expert Systems with Applications 36, no. 9 (2009): 11509-11516.

[13] Ohzeki, Kazuo, and Masaru Sakurai, "SVD-based watermark with quasi-one-way operation by reducing a singular value matrix rank", In Proceedings of the 1st international conference on Forensic applications and techniques in telecommunications, information, and multimedia and workshop, p. 21. ICST, 2008.

[14] Liu, Ruizhen, and Tieniu Tan. "An SVD-based watermarking scheme for protecting rightful ownership", Multimedia, IEEE Transactions on 4, no. 1 (2002): 121-128.

[15] Yan, Hongshou, and Weimin Yang. "A Watermarking Algorithm Based on Wavelet and Hadamard Transform for Color Image." Journal of Software Engineering and Applications 6 (2013): 58.

[16] Ahmed, Kamal A., Hussain Al-Ahmad, and Patrick Gaydecki, "Robust image watermarking using two dimensional Walsh coding." In Image Processing (IPR 2012), IET Conference on, pp. 1-5. IET, 2012.

[17] Jiansheng, Mei, Li Sukang, and Tan Xiaomei. "A digital watermarking algorithm based on DCT and DWT", In Proceedings of the International Symposium on Web Information Systems and Applications, pp. 104-107, 2009.

[18] Amirgholipour, Saeed K., and Ahmad Reza Naghsh-Nilchi. "Robust Digital Image Watermarking Based on Joint DWT-DCT." JDCTA 3, no. 2 (2009): 42-54.

[19] Sharkas, Maha, Dahlia ElShafie, and Nadder Hamdy. "A Dual Digital-Image Watermarking Technique." In WEC (5), pp. 136-139. 2005.

[20] Singh, Surya Pratap, Paresh Rawat, and Sudhir Agrawal, "A robust watermarking approach using DCT-DWT", International Journal of Emerging Technology and Advanced Engineering, Volume 2, Issue 8, 2012.

[21] Kekre, H. B., Tanuja Sarode, and Shachi Natu. "Performance Improvement by Sorting the Transform Coefficients of Host and Watermark using Unitary Orthogonal Transforms Haar, Walsh and DCT", Advances in Image and Video Processing 2, no. 4 (2014): 24-38.

[22] H. B. Kekre, Archana A. Athawale, Dipali Sadavarti, “Algorithm to Generate Kekre’s Wavelet Transform from Kekre’s Transform”, International Journal of Engineering, Science and Technology, (IJEST) Vol. 2(5), 756-767, 2010.

[23] Kekre, H. B., and Tanuja K. Sarode. "Fast Codebook Generation Algorithm for Color Images using Vector Quantization." International Journal of Computer Science and Information Technology 1.1 (2009): 7-12

[24] Dr. H. B. Kekre, Dr. Tanuja Sarode, Shachi Natu, “Image Zooming using Sinusoidal Transforms like Hartley, DFT, DCT, DST and Real Fourier Transform”, International journal of computer science and information security Vol. 12 No. 7, pp. 11-16, July 2014.

[25] H. B. Kekre, Tanuja Sarode, Sudeep Thepade, “Grid based image scaling technique”, International Journal of Computer Science and Applications, Volume 1, No. 2, pp. 95-98, August 2008


www.ijcsit.com 313

68. effect of varying embedding energy and energy wise ... 6/vol6issue01/ijcsit2015060168.pdf ·...

Documents